Incommensurability

In the [YOU] on AI Field Guide

The cycle documents the camps that form, the positions that calcify, the frustration of a discourse that cannot resolve itself through better arguments or more evidence. Incommensurability explains why. When a senior developer insists that real programming requires machine-level understanding, and a natural-language builder insists that directing an AI tool to produce a working result is real programming, neither is making an empirical claim. Both are expressing paradigmatic commitments that assign different meanings to the same word. The senior developer's evidential standard—quality of the code, elegance of the architecture, depth of the developer's comprehension—is invisible from the new paradigm, where those markers are legacy criteria applied without justification to a framework in which the human's task is specifying intention, not understanding implementation. The new paradigm's evidential standard—does the product work, was it built efficiently—looks like shallowness from the old paradigm.

The incommensurability extends to the perception of the same artifacts. An old-paradigm practitioner looking at AI-generated code sees absence: the absence of idiomatic patterns that signal deep language knowledge, the absence of architectural choices that reflect system-level thinking. A new-paradigm practitioner looking at the same code sees presence: working functionality, speed, a result that enables the next decision. They are not looking at different things. They are looking at the same thing through different perceptual frameworks, and Kuhn showed that this kind of theory-laden perception is not correctable by pointing at the artifact more carefully.

The practical consequence in organizations is friction that resists the usual mechanisms of resolution. Meetings, documentation, and agreed-upon standards assume that the parties share the conceptual vocabulary required for deliberation. When the parties are incommensurable, the vocabulary of deliberation is itself contested. The senior developer who reviews AI-generated code and flags it as “not up to standard” is applying old-paradigm standards to new-paradigm output. The junior developer who defends it as “good enough” is applying new-paradigm standards the senior does not recognize as legitimate. The manager who mediates discovers that the conflict cannot be resolved by splitting the difference, because the difference is not about degree but about kind.

Origin

Kuhn introduced the concept in 1962 with less precision than his subsequent defenders and critics sometimes attribute to him. The first edition of Structure stated the incommensurability thesis in ways that invited the reading that paradigm-choice was irrational—that if one cannot translate between paradigms, then there is no rational basis for preferring one over another. This reading made Kuhn appear to be a relativist about science, which provoked strong negative reactions from philosophers of science who saw him as undermining the rationality of scientific progress. Kuhn spent the next three decades clarifying. In a 1982 paper, “Commensurability, Comparability, Communicability,” he argued that incommensurability does not imply total failure to communicate or irrational paradigm choice. What it implies is that the translation between paradigms is never complete: something is always lost, a remainder that cannot be carried across, and the choice between paradigms involves judgment, values, and the assessment of potential that cannot be reduced to algorithmic application of a neutral criterion.

The concept has been taken up in science studies, in the history of economics, in comparative literature, and, increasingly, in the study of technological transitions. Giovanni Dosi's 1982 paper on technological paradigms extended Kuhn's framework to technology directly, arguing that “technological paradigms” define relevant problems and search heuristics in exactly the way Kuhn's scientific paradigms do—with the corollary that transitions between technological paradigms produce the same incommensurability, the same mutual incomprehension between practitioners who have organized their competence around different frameworks.

Key Ideas

Meaning change across paradigm boundaries. Key terms do not merely refer to different things in different paradigms; they carry different conceptual structures, different networks of implication, different standards of application. “Programming,” “expertise,” “understanding,” and “real work” all exhibit this drift in the current AI transition. The drift is invisible until the conversation turns to foundations, at which point the parties discover they have been talking across a conceptual gap.

Theory-laden perception. Paradigms function as perceptual filters. The same artifact, the same code, the same output looks different depending on which paradigm is doing the looking. This is not a matter of different interpretations of shared data; it is a matter of what counts as data and what counts as significance. Old-paradigm practitioners looking at AI-generated code see the absence of markers of genuine competence. New-paradigm practitioners see the presence of functional results. Both are perceiving something real. Neither perception is available to the other without a change in paradigm.

Resolution through replacement, not persuasion. Kuhn argued that paradigm conflicts are not resolved through better arguments, because the arguments on both sides are internally coherent and the evidential standards by which arguments are evaluated are themselves paradigm-relative. Resolution comes through the social processes by which one paradigm gains adherents: some practitioners convert, others retire, new practitioners are trained who know no other framework. The AI transition is compressing this process into months rather than decades, with consequences for practitioners that Kuhn's historical method could only partially anticipate.

The productive remainder. What is lost in translation across paradigm boundaries is not simply noise; it is the knowledge that each paradigm had developed that the other paradigm's framework cannot accommodate. The old paradigm's deep knowledge of implementation—how machines actually work, why systems fail in particular ways, what elegant architecture looks like—is not refuted by the new paradigm. It is rendered less central. But it is the precisely this knowledge that the new paradigm most needs when its tools fail in ways their directions did not anticipate. The practitioners who survive the transition usefully are those who can access both paradigms simultaneously.